Abstract
Predicting customer churn is an essential part of retention strategy for telecom companies so as to maximize revenue. In this paper, four machine learning models, Random Forest, Gradient Boosting, Logistic Regression, and K-Nearest Neighbors are compared to predict customer churn using a telecom dataset. We use SMOTE-Tomek to cope with class imbalance and optimize models by using GridSearchCV, Optuna, and Grey Wolf Optimizer. Our optimized Random Forest has 85.9% of accuracy beating other models. The study reveals the main churn factors such as type of contract and the usage of services, which are useful in developing targeted retention strategies for telecom providers..
